COMPAS and Algorithmic Risk Assessment Tools in U.S. Sentencing

Algorithmic risk assessment tools occupy a contested position in the U.S. criminal justice system, where they influence decisions ranging from pretrial detention to sentencing and parole. COMPAS (Correctional Offender Management Profiling for Alternative Sanctions), developed by Equivant (formerly Northpointe), is the most litigated and studied of these instruments. This page provides a reference-grade examination of how these tools work, what legal challenges they have generated, where classification boundaries fall, and what tradeoffs courts and legislatures continue to debate.

• Definition and Scope
• Core Mechanics or Structure
• Causal Relationships or Drivers
• Classification Boundaries
• Tradeoffs and Tensions
• Common Misconceptions
• Checklist or Steps (Non-Advisory)
• Reference Table or Matrix
• References

Definition and Scope

Algorithmic risk assessment tools in criminal justice are structured instruments that generate a numerical score or risk tier predicting the likelihood that an individual will reoffend, fail to appear at trial, or violate conditions of release. These tools are used across the criminal justice pipeline — at pretrial detention hearings, at sentencing, in parole and probation determinations, and in corrections classification.

COMPAS produces scores across multiple domains, including a General Recidivism Risk score, a Violent Recidivism Risk score, and a Pretrial Release Risk score. Each score is expressed on a scale of 1 to 10, with 10 representing the highest assessed risk. The instrument draws on 137 input items covering criminal history, age at first arrest, residential instability, and responses to attitudinal surveys (Northpointe/Equivant, COMPAS Risk and Need Assessment System User Guide).

The scope of deployment extends well beyond COMPAS. The Arnold Foundation's Public Safety Assessment (PSA), the Ohio Risk Assessment System (ORAS), the Level of Service Inventory–Revised (LSI-R), and the Virginia Pretrial Risk Assessment Instrument (VPRAI) represent a broader ecosystem of tools embedded in jurisdictions across all 50 states. As of reporting by the Pretrial Justice Institute, pretrial risk assessment tools operate in jurisdictions covering the majority of the U.S. population, though no single federal mandate requires their use.

The legal framework governing these tools is diffuse. The U.S. Supreme Court addressed algorithmic sentencing instruments indirectly in Loomis v. Wisconsin (2016), where the Wisconsin Supreme Court upheld COMPAS use at sentencing, reasoning that the tool was one of multiple factors considered. The U.S. Supreme Court declined to hear the case in 2017. Constitutional challenges have proceeded under the Due Process Clause of the Fourteenth Amendment, the Sixth Amendment right to confront adverse evidence, and the Eighth Amendment prohibition on arbitrary punishment. The algorithmic due process framework emerging from these cases has no settled federal resolution.

Core Mechanics or Structure

COMPAS operates as a proprietary actuarial instrument. The algorithm ingests structured inputs — primarily interview responses and criminal history records — and applies a weighted scoring model to produce risk tier assignments. The precise weights assigned to each variable are treated as trade secrets by Equivant, which is the central source of the transparency dispute.

The scoring model is actuarial, not clinical. Rather than a trained clinician rendering an individual judgment, the instrument compares an individual's profile to statistical patterns derived from a reference population. The reference population used in COMPAS validation studies was drawn from Broward County, Florida, and Tuscola County, Michigan, populations — a methodological choice that has direct implications for cross-jurisdictional validity (ProPublica, "Machine Bias," 2016).

The tool's input variables span five domains:
1. Criminal involvement — prior arrests, charges, convictions, juvenile record
2. Relationships and lifestyle — family criminal history, peer associations
3. Personality and attitudes — self-reported measures of anger, criminal thinking, social isolation
4. Residential instability and financial — housing history, employment history
5. Substance abuse — self-reported and documented use history

Race is not a direct input variable. However, variables correlated with race — particularly those capturing socioeconomic conditions such as neighborhood stability, employment, and family criminal history — function as proxies. The AI bias in criminal justice analysis of COMPAS has consistently focused on this proxy-variable mechanism.

Causal Relationships or Drivers

The adoption of algorithmic risk assessment tools was driven by four converging forces. First, the 1984 Sentencing Reform Act and the creation of the U.S. Sentencing Commission established a norm of structured, evidence-based sentencing intended to reduce judicial disparity. Second, the bail reform movement — accelerating after the 2010s — sought to replace cash bail with risk-based release decisions, which created institutional demand for standardized instruments. Third, correctional systems under fiscal pressure needed classification tools to allocate supervision and programming resources efficiently. Fourth, the broader movement toward evidence-based practices in criminal justice, formalized through the Justice Reinvestment Initiative supported by the Bureau of Justice Assistance (Bureau of Justice Assistance, Justice Reinvestment Initiative), provided funding incentives for jurisdictions adopting validated instruments.

The ProPublica investigation published in May 2016 demonstrated — using 7,000 Broward County defendants tracked over 2 years — that COMPAS scores were racially disparate in predictive error patterns. Black defendants who did not reoffend were labeled higher risk at nearly twice the rate of white defendants who did not reoffend (45% vs. 24%). White defendants who did reoffend were labeled lower risk at a disproportionate rate (ProPublica, "Machine Bias," 2016). Equivant disputed the analysis methodology, and subsequent academic debate centered on whether predictive parity or calibration — two mathematically incompatible fairness criteria — should serve as the operative standard.

The AI sentencing guidelines landscape reflects this causal tension: demand for consistency and fiscal efficiency pulling toward algorithmic tools, and constitutional equality principles pulling toward scrutiny of their outputs.

Classification Boundaries

Algorithmic risk tools in U.S. criminal justice can be classified along four axes:

By decision point:
- Pretrial (PSA, VPRAI, COMPAS Pretrial) — governs release conditions
- Sentencing (COMPAS General/Violent, LSI-R) — informs incarceration length and conditions
- Corrections/classification (COMPAS Institutional) — governs internal housing and programming
- Parole/reentry (ORAS, LSI-R) — governs supervised release conditions

By methodological type:
- Actuarial (statistical, group-based) — COMPAS, LSI-R, PSA
- Clinical (individual professional judgment) — court-ordered psychological evaluations
- Hybrid structured professional judgment — tools that combine scoring with clinician override capacity

By proprietary status:
- Proprietary/closed-source — COMPAS, LSI-R (Multi-Health Systems)
- Open-source or publicly documented — PSA (Arnold Foundation, methodology published), ORAS (Ohio Department of Rehabilitation and Correction, publicly available)

By federal vs. state deployment:
- Federal pretrial: the 18 U.S.C. § 3142 Bail Reform Act framework does not mandate any specific tool; the Administrative Office of the U.S. Courts has developed the Pretrial Risk Assessment (PTRA) instrument for federal use
- State sentencing: at least 20 states had incorporated risk assessment into sentencing statute or practice as of the National Conference of State Legislatures' 2021 tracking, though specific counts evolve with legislation

Tradeoffs and Tensions

The core tension in algorithmic sentencing is between two legitimate values: consistency and individualization. Structured tools reduce the documented disparity produced by unconstrained judicial discretion — a problem extensively documented in the U.S. Sentencing Commission's reports on inter-judge variation. At the same time, actuarial tools score individuals based on group-level statistics, which raises Eighth Amendment concerns about punishing people for characteristics they share with others rather than for their own conduct.

The mathematical fairness impossibility theorem — formalized by researchers Chouldechova (2017, Fair Prediction with Disparate Impact, published in Big Data journal) and Kleinberg et al. — demonstrates that when base rates of recidivism differ across demographic groups (as they do in U.S. data), no single tool can simultaneously achieve calibration (equal accuracy across groups), equal false positive rates, and equal false negative rates. Any jurisdiction choosing a risk tool is implicitly choosing which fairness criterion to optimize, a choice that has never been made explicit in any U.S. jurisdiction's statutory framework.

The transparency-accuracy tradeoff is equally unresolved. Proprietary tools may perform better due to proprietary training data, but defendants cannot inspect or challenge the model. In State v. Loomis, the Wisconsin Supreme Court held that because COMPAS was not the determinative factor and a pre-sentence investigation report was available, due process was not violated — but this holding has not been adopted universally and applies only to Wisconsin's specific procedural context.

The AI pretrial detention debate adds a distinct dimension: pretrial risk scores affect individuals who have not been convicted of any offense, making the constitutional stakes of error higher than at post-conviction sentencing.

Common Misconceptions

Misconception: COMPAS determines sentences.
Correction: No U.S. jurisdiction uses COMPAS as a determinative sentencing mechanism. Courts have consistently held, including in Loomis, that a judge must exercise independent discretion and that the tool constitutes one informational input among others documented in a presentence investigation report.

Misconception: COMPAS uses race as an input variable.
Correction: Race is not a direct input. The disparity identified by ProPublica operates through proxy variables correlated with race — criminal history length, socioeconomic factors, residential instability — which are direct inputs. This distinction is legally and technically significant.

Misconception: Higher COMPAS accuracy means fairness.
Correction: Overall predictive accuracy (AUC) and group-level fairness are independent properties. Equivant reported AUC scores of approximately 0.70 for general recidivism prediction, which is statistically moderate, but that aggregate figure does not address differential error rates across demographic groups — the central fairness concern.

Misconception: Open-source tools solve the bias problem.
Correction: Transparency enables independent auditing but does not eliminate bias embedded in training data. The Arnold Foundation's PSA is publicly documented, yet researchers have identified disparate impacts in its application in jurisdictions including New Jersey (Laura and John Arnold Foundation, PSA documentation).

Misconception: Federal courts widely use COMPAS.
Correction: Federal courts use the Administrative Office of the U.S. Courts' PTRA instrument for pretrial decisions. COMPAS is primarily a state-court and corrections tool. The federal sentencing framework under the U.S. Sentencing Commission guidelines (USSC) does not reference COMPAS.

Checklist or Steps (Non-Advisory)

The following sequence describes the procedural stages through which a risk assessment score typically moves in a state criminal case where such a tool is employed:

1. Intake and data collection — Correctional or pretrial staff collect structured interview responses and pull criminal history records from state and federal databases (e.g., NCIC).
2. Instrument administration — The designated tool (COMPAS, PSA, ORAS, etc.) processes the inputs; a risk tier and/or numeric score is generated.
3. Score integration into case file — The score is incorporated into a presentence investigation report (PSI/PSR) prepared by probation officers, or into a pretrial services report for bail hearings.
4. Judicial receipt — The sentencing judge or magistrate receives the report containing the score as part of the case record.
5. Defense notification — In jurisdictions with disclosure requirements, defense counsel receives the report, including the score; some jurisdictions also provide access to the response data used to generate the score.
6. Challenge opportunity — Defense counsel may dispute input accuracy (e.g., incorrect criminal history records), contest the score's applicability, or present expert testimony on methodological limitations; the procedural vehicle varies by jurisdiction.
7. Judicial weighing — The judge incorporates the score as one of multiple enumerated factors under state sentencing statutes; the judge documents the basis for the sentence.
8. Appellate record — The PSR, including the risk score, becomes part of the trial court record and is available for appellate review of the sentence.
9. Post-sentence corrections use — The score may be updated and used for internal classification, programming assignment, and parole eligibility calculations.

Reference Table or Matrix

Tools classified as "No (public)" have published methodology documentation; "No (internal)" indicates government-operated instruments with partial public documentation.

The AI in federal courts context for the PTRA differs materially from COMPAS use in state systems: federal pretrial services officers administer the PTRA under the supervision of the Administrative Office of the U.S. Courts, and outcomes are governed by the Bail Reform Act's enumerated factors rather than a state sentencing grid.

For a broader view of how algorithmic tools intersect with parole and probation decisions, including post-release supervision scoring, the classification distinctions in this table are the operative starting point.

References

• Equivant / Northpointe — COMPAS Risk and Need Assessment System (Wisconsin DOC documentation)
• ProPublica — "Machine Bias: Risk Assessments in Criminal Sentencing" (2016)
• Arnold Foundation — Public Safety Assessment: Tool Documentation
• U.S. Sentencing Commission (USSC)
• Bureau of Justice Assistance — Justice Reinvestment Initiative
• Administrative Office of the U.S. Courts — Pretrial Services
• National Conference of State Legislatures — Risk and Needs Assessment in the Criminal Justice System
• Pretrial Justice Institute
• Ohio Department of Rehabilitation and Correction — ORAS Documentation
• [U.S. Parole Commission — Salient Factor Score, 28 C.F.R. Part 2](https://www